Merge pull request #104 from flatironinstitute/cufinufft-request-devi…

…ce-id

Fix: Automatically set cufinufft's gpu_device_id parameter

.pre-commit-config.yaml

@@ -17,7 +17,7 @@ repos:
 
   # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
   - repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 23.9.1
+    rev: 24.1.1
     hooks:
       - id: black
 

examples/convolution_2d.py

@@ -3,7 +3,6 @@
 =================
 """
 
-
 #######################################################################################
 # Import packages
 # ---------------

pytorch_finufft/functional.py

@@ -2,6 +2,7 @@
 Implementations of the corresponding Autograd functions
 """
 
+import functools
 import warnings
 from typing import Any, Callable, Dict, Optional, Tuple, Union
 
@@ -36,12 +37,16 @@
 
 
 def get_nufft_func(
-    dim: int, nufft_type: int, device_type: str
+    dim: int, nufft_type: int, device: torch.device
 ) -> Callable[..., torch.Tensor]:
-    if device_type == "cuda":
+    if device.type == "cuda":
         if not CUFINUFFT_AVAIL:
             raise RuntimeError("CUDA device requested but cufinufft failed to import")
-        return getattr(cufinufft, f"nufft{dim}d{nufft_type}")  # type: ignore
+        # note: in the future, cufinufft may figure out gpu_device_id on its own
+        # see: https://github.com/flatironinstitute/finufft/issues/420
+        return functools.partial(
+            getattr(cufinufft, f"nufft{dim}d{nufft_type}"), gpu_device_id=device.index
+        )
 
     if not FINUFFT_AVAIL:
         raise RuntimeError("CPU device requested but finufft failed to import")
@@ -137,7 +142,7 @@ def forward(  # type: ignore[override]
         # pop because cufinufft doesn't support modeord
         modeord = finufftkwargs.pop("modeord", FinufftType1.MODEORD_DEFAULT)
 
-        nufft_func = get_nufft_func(ndim, 1, points.device.type)
+        nufft_func = get_nufft_func(ndim, 1, points.device)
 
         batch_dims = values.shape[:-1]
         finufft_out = nufft_func(
@@ -217,7 +222,7 @@ def backward(  # type: ignore[override]
         grads_points = None
         grad_values = None
 
-        nufft_func = get_nufft_func(ndim, 2, device.type)
+        nufft_func = get_nufft_func(ndim, 2, device)
 
         if any(ctx.needs_input_grad):
             if _mode_ordering:
@@ -317,7 +322,7 @@ def forward(  # type: ignore[override]
         if modeord:
             targets = batch_fftshift(targets, ndim)
 
-        nufft_func = get_nufft_func(ndim, 2, points.device.type)
+        nufft_func = get_nufft_func(ndim, 2, points.device)
         batch_dims = targets.shape[:-ndim]
         shape = targets.shape[-ndim:]
         finufft_out = nufft_func(
@@ -378,7 +383,13 @@ def vmap(  # type: ignore[override]
     @staticmethod
     def backward(  # type: ignore[override]
         ctx: Any, grad_output: torch.Tensor
-    ) -> Tuple[Union[torch.Tensor, None], Union[torch.Tensor, None], None, None, None,]:
+    ) -> Tuple[
+        Union[torch.Tensor, None],
+        Union[torch.Tensor, None],
+        None,
+        None,
+        None,
+    ]:
         _i_sign = ctx.isign
         _mode_ordering = ctx.mode_ordering
         finufftkwargs = ctx.finufftkwargs
@@ -404,7 +415,7 @@ def backward(  # type: ignore[override]
 
         if ctx.needs_input_grad[0]:
             # wrt. points
-            nufft_func = get_nufft_func(ndim, 2, points.device.type)
+            nufft_func = get_nufft_func(ndim, 2, points.device)
 
             coord_ramps = coordinate_ramps(shape, device)
 
@@ -422,7 +433,7 @@ def backward(  # type: ignore[override]
 
         if ctx.needs_input_grad[1]:
             # wrt. targets
-            nufft_func = get_nufft_func(ndim, 1, points.device.type)
+            nufft_func = get_nufft_func(ndim, 1, points.device)
 
             grad_targets = nufft_func(
                 *points,

tests/test_t1_forward.py

@@ -69,3 +69,9 @@ def test_t1_forward_CPU(N, dim) -> None:
 @pytest.mark.parametrize("N, dim", Ns_and_dims)
 def test_t1_forward_cuda(N, dim) -> None:
     check_t1_forward(N, dim, "cuda")
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="require multiple GPUs")
+def test_t1_forward_cuda_device_1() -> None:
+    # added after https://github.com/flatironinstitute/pytorch-finufft/issues/103
+    check_t1_forward(3, 1, "cuda:1")